Annotation efficient learning with affinity graphs

Thesis

Annotation-efficient learning has emerged as a critical area of research due to the scarcity of labeled samples posing a substantial barrier to developing robust fullysupervised deep neural networks. When the availability of labeled samples is limited, developing an effective learning system becomes a formidable challenge. Conversely, unlabeled data is often plentiful and can be obtained at a relatively low cost. Consequently, the concept of leveraging a substantial volume of unlabeled data to train deep models, despite the paucity of labeled samples, emerges as a compelling proposition.

This thesis explores novel approaches to annotation-efficient learning in machine learning, with a focus on leveraging implicit relationships within data to improve model performance in scenarios with limited labeled data. The research addresses three key challenges: effectively utilizing implicit structures in input data, integrating these structures into the learning process, and determining optimal representations for extracting and utilizing implicit relationships. The methodology centers on the development and application of affinity graph constraints across three domains: self-supervised learning for whole-slide image analysis, semi-supervised learning for medical image segmentation, and multi-modal learning for single-cell data integration.

The methodology centers on the development and application of affinity graph constraints across three domains: self-supervised learning for whole-slide image analysis, semi-supervised learning for medical image segmentation, and multi-modal learning for single-cell data integration. Specifically, we propose:
1. An affinity graph constraint (AGC) for self-supervised learning on whole-slide images, which captures fine-grained features and improve existing self-supervised methods.
2. An affinity-graph-guided contrastive learning framework for semi-supervised medical image segmentation, incorporating patch-wise class-centric sampling and hard-negative reweighting.
3. A single-cell Affinity Graph transFormer (scAGFormer) for multi-modal singlecell analysis, which employs an affinity graph prior to improve modality transformation.

Key findings demonstrate significant improvements over state-of-the-art methods across all three domains. The proposed affinity graph-based approaches consistently enhance model performance, particularly in scenarios with limited labeled data. The meworks show remarkable generalizability across different datasets and tasks, highhting their potential for broad application in medical image analysis and computational biology.

Authors: Cheng, Z

• DOI: 10.5287/ora-44aj4d9qe
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: annotation-efficient learning; medical-imaging analysis; machine learning
• Subjects: Semi-supervised Learning (Machine Learning); Deep learning (Machine learning)